|
|
| 基于态势Token表征与多智能体Transformer的多无人机协同避障方法 |
| Multi-UAV Cooperative Obstacle Avoidance Based on Situation-Token Representation and Multi-Agent Transformer |
| 投稿时间：2026-07-28 修订日期：2026-08-08 |
| DOI： |
| 中文关键词: 无人机 协同避障 态势Token 多智能体Transformer 强化学习 |
| English Keywords:UAV cooperative obstacle avoidance situation token multi-agent Transformer reinforcement learning |
| 基金项目:国家自然科学基金资助项目 |
|
| 摘要点击次数: 139 |
| 全文下载次数: 0 |
| 中文摘要: |
| 针对多无人机在密集障碍环境中协同避障时，不同类型的局部观测难以有效融合、LiDAR测距信息的空间特征利用不足以及多机动作难以协调的问题，提出一种基于态势Token表征与多智能体Transformer的多无人机协同避障方法。该方法采用卷积神经网络提取LiDAR空间特征，简称CNN Token-MAT。该方法以单架无人机为基本单元，将本机状态、目标信息、LiDAR感知和邻机状态融合为态势Token，并利用多智能体Transformer完成跨机态势编码与多机动作生成。仿真实验结果表明，在无人机数量N=4、静态障碍物数量为60～120的条件下，CNN Token-MAT的到达率均高于Flat-MAT，其中最高提高14.95个百分点。结果验证了所提方法在四无人机密集静态障碍环境中的有效性，同时表明卷积式LiDAR编码能够进一步改善多无人机协同避障性能。 |
| English Summary: |
| To address the difficulty of effectively fusing different types of local observations, insufficient use of spatial features in LiDAR range measurements, and poor coordination of multi-UAV actions during cooperative obstacle avoidance in dense-obstacle environments, a multi-UAV cooperative obstacle avoidance method based on situation-token representation and a mul-ti-agent Transformer is proposed. The method employs a convolutional neural network to extract LiDAR spatial features and is termed CNN Token-MAT. Taking each UAV as the basic unit, the method integrates the UAV state, target information, LiDAR perception, and neighboring-UAV states into a situation token, and uses the multi-agent Transformer for inter-UAV situation en-coding and multi-UAV action generation. Simulation results with N=4 UAVs and 60-120 static obstacles show that CNN Token-MAT consistently achieves higher arrival rates than Flat-MAT, with a maximum improvement of 14.95 percentage points. The results validate the effectiveness of the proposed method in a dense static-obstacle environment with four UAVs and indicate that convolutional LiDAR encoding further improves multi-UAV cooperative obstacle avoidance performance. |
| 查看/发表评论 下载PDF阅读器 |
